Product Craft, Growth, and AI Strategy

Summary

Conversation with product leader and investor Peter Deng on building and scaling iconic products, AI’s impact, hiring, management, and product craft, drawing lessons from Facebook, Instagram, Uber, OpenAI, Airtable, and Oculus.

Action Items

• Ongoing – Founders: Clarify your unique data flywheel and workflow advantage when building on top of LLMs; codify this in a clear thesis.
• Ongoing – Product leaders: Audit instrumentation and logging; ensure you can accurately measure retention and key funnel metrics before pushing growth.
• Ongoing – Managers: Define and share your own “API” (how to work with me) with direct reports, including expectations on autonomy and feedback.
• Ongoing – PMs/builders: Systematically dogfood your product (including in “extreme” real contexts) to deeply empathize with users and uncover gaps.
• Ongoing – Job seekers: Evaluate opportunities by learning potential, mission alignment with human behavior, and clarity of founders’ unique insight.
• Ongoing – Founders/PMs: Explicitly assign “growth” and “craft” ownership to different leaders to create healthy tension between metrics and product taste.

Building Products & Scaling from 0→1→100

• Moving from 0→1 to 1→100

  • 0→1 is finding product-market fit; 1→100 is orchestrating hyperscale.
  • In 1→100, you must “plan your chess moves” in advance and build systems that let you go sustainably faster.
  • Sometimes you must go slow (architecture, systems, infra) so later you can go very fast.

• Systems thinking examples

  • Facebook News Feed (current architecture):
    • Carefully designed end-to-end sharing loop: post → feed exposure → likes → notifications → repeat.
    • Information architecture built to last; minimal structural change over ~12 years.
  • Uber Rider app:
    • Re-architected messy “spaghetti” code into scalable abstractions like “venues” for pickup/dropoff.
    • Venue abstraction enabled generalized solutions for airports, complex locations, and international markets.
  • Messenger:
    • Deep investment in infra (e.g., notifications) helped scale to ~4.7B messages/day in ~2.5 years.

• Measurement and growth discipline

  • Do not “fly the plane without instruments” – measure everything.
  • Early growth teams expose missing logging, weak rigor, and unclear metrics.
  • Growth PMs force deeper analysis, hypothesis-driven experiments, and a culture of data-informed decisions.
  • Retention (cohorted, asymptotic curves) is the key indicator for product viability, not just topline usage.

• Portfolio approach to product work

  • Think in portfolios (not binary “we scale now vs later”):
    • For mature companies: 70/20/10 (core/adjacent/bets) can make sense.
    • For startups: portfolio mix may be closer to 50/50 between scaling and new bets.
  • Adjust mix based on stage and product maturity.

What Really Matters in Products

• When product “doesn’t matter” (pixels vs reality)

  • At Uber, the true product for riders was price and ETA, not just UI polish.
  • Fixing UI bugs often had less impact than improving marketplace dynamics, operations, and reliability.
  • Many great tech companies are operations or business model companies wrapped in tech (e.g., Uber).

• Tech breakthroughs vs execution

  • Many valuable companies did not start from proprietary tech breakthroughs:
    • Facebook built on basic databases and connections, then added products like News Feed and tagging.
    • Uber leveraged existing GPS and smartphones plus marketplace design and ops.
  • Huge value often comes from “elbow grease” and connecting obvious-seeming dots, not lab-only innovation.
  • Even when you have a tech breakthrough, product experience, ergonomics, and workflows rapidly become decisive.

• Craft vs metrics – keeping tension healthy

  • You need explicit tension between growth (metrics) and craft (experience, aesthetics).
  • Assign different leaders to own each dimension so both are advocated strongly.
  • Leader’s role is to adjudicate tradeoffs and stretch the product across the full spectrum.

Types of Product Managers

• Everyone usually has a primary and secondary archetype.
• Hiring: build an “Avengers team” where spikes are complementary across these archetypes.

AI, AGI, and the Future of Products

• AGI realism and optimism

  • AGI is necessary but not sufficient for solving all problems.
  • Huge value will still require “hustle and elbow grease” from builders to channel this new energy.
  • Every major technology triggered early fear (e.g., bicycles once thought catastrophic) before being humanized.
  • Humans and technology co-evolve: fear → familiarity → mastery as people adapt and build new systems.

• Timeline attitudes

  • Early ChatGPT era: strong fear; 18 months later: more familiarity, flourishing of startups and products.
  • Expect similar attitude shifts as capabilities progress.

• Large model companies & product’s role

  • Big leverage comes from tight collaboration between product and post-training / research teams.
  • At OpenAI and Anthropic: PMs embedded with researchers can influence model behavior, alignment, and capabilities.
  • Over time, advantage shifts from “raw intelligence” to:
    • Fine-tuning models to user needs and workflows.
    • Designing model behavior and persona (model design).

• Model design as a function

  • “Model designers” blend technical depth and product taste to shape how models behave.
  • Role codified at OpenAI (e.g., led by Joanne Jen) to own model vibe, usability, and alignment.

Building AI Startups on LLMs

• You can begin without proprietary data if your workflow lets you accumulate distinctive usage data (e.g., Windsurf using foundational models, then training on accept/reject signals).
• Distribution moats (e.g., Microsoft, Google, Zoom) can be overcome if your product is dramatically better in craft and fit (examples: Cursor, Windsurf, Lovable, Bolt, Granola).

Education & Language in the Age of AI

• AI and education

  • Exposure to AI changes how kids think and “rewires” their brains.
  • Example: Peter’s 9-year-old son built a custom GPT that, for any topic, generates a pangram sentence containing every letter of the alphabet.
  • This kind of prompting is nearly impossible in traditional programming, but natural with LLMs.
  • Programming will likely become automated, but learning to program remains valuable for structured thinking.

• Skills that will matter more

  • Asking good questions and formulating effective prompts will be a major differentiator.
  • As calculators made rote arithmetic less essential and Google made memorization less critical, AI will automate code; humans must move up the abstraction stack.
  • Future education should focus on:
    • Higher-level abstraction and reasoning.
    • Creativity and curiosity.
    • Framing problems and questions well.

• Instructors and AI

  • Professors already use ChatGPT to create curricula; students sometimes negatively rate them when they detect AI usage.
  • Indicates that entire educational systems, not only student behavior, must adapt.

• Power of language and thought

  • Peter was influenced by a “Language and Thought” class (Herbert Clark): language shapes how we think.
  • Bilingual examples and studies (e.g., different words for shades of blue) suggest language changes perception and categorization.
  • In work:
    • He obsessively crafts words in decks and docs (e.g., 20-word decks refined over hours).
    • Poor wording in vision docs or PRDs causes misalignment and misinterpretation.
  • Poetic fit: AI breakthroughs came from large language models and Python “language” execution; human knowledge is heavily encoded in language itself.

Management, Hiring & Career Growth

Building Teams Like Products

• Treat the team as a product

  • Design team composition intentionally; do not just fill “warm bodies.”
  • Aim for a “team of Avengers”: each member spikes in a distinct dimension (growth, craft, infra, business, research).
  • Leader’s job: create healthy debates and adjudicate tradeoffs.

• Distribution of strengths

  • Look at what your company truly needs (stage, market, product gaps).
  • Hire specific archetypes to fill those needs (e.g., growth PM, platform PM, model designer).

Two Core Hiring Principles

• Principle 1: 6‑month autonomy test

  • Peter’s rule: “In 6 months, if I’m telling you what to do, I’ve hired the wrong person.”
  • Effects:
    • Forces high bar in hiring; avoid settling.
    • Sets clear expectations for candidates and reports about desired autonomy.
    • Shifts the meta-goal from “hit this OKR” to “are we calibrating so that in 6 months you are telling me what needs to be done?”
    • Works for all managers, not just executives; helps scale leadership and institutional knowledge.

• Principle 2: Growth mindset as the non‑negotiable

  • Growth mindset is Peter’s top hiring filter.
  • Without it, feedback and development stall; no meta-learning occurs.
  • Final interview (as CPO/head of product) focuses almost entirely on growth mindset, trusting others to evaluate product sense, design, metrics, etc.

• Growth mindset interview question

  • Ask:
    • “Tell me about one of the biggest mistakes you’ve made. The more painful, the better.”
    • “What exactly happened? How did it change how you think and work now?”
  • Look for:
    • Genuine vulnerability vs defensive/PR answers.
    • Depth of reflection: do they have a clear, behavior-changing takeaway?
    • Ability to convert loss into lesson.
  • Secondary benefit: builds a foundation of psychological safety and candor if they join.

“PXD API”: How to Work with Peter

• Peter maintains a “PXD API” doc (how to work with me), including:
  • The 6‑month autonomy expectation.
  • Traits of people who thrive working with him (e.g., growth mindset, feedback-seeking).
  • Clear up-front contract fosters alignment and transparency.

Operating Principle: Say–Do–Say

• Managing up and operating effectively:
  • Say you’re going to do the thing.
    • Align on goals; use precise language; invite correction if priorities shifted.
  • Say that you’re doing the thing.
    • Provide progress updates; reconfirm importance; surface if course-correction is needed.
  • Say that you did the thing.
    • Close the loop; make impact visible; avoid “invisible work,” especially for introverts.
  • Applies to managing up, down, and across; also mirrors classic presentation advice: tell them what you’ll say, say it, then recap.

Leaning into Strengths

• Fit is a two-way street

  • People should not force themselves into archetypes they don’t resonate with.
  • Managers should help reports articulate their strengths and passions (often via writing) and adjust roles accordingly.
  • Example:
    • Joanne Jen at OpenAI had rare depth in both tech and taste.
    • Peter nudged her to write down what she loved doing → codified “model designer” role → hired others into the function → major impact on ChatGPT’s model vibe and usability.

• Career advice

  • Life is short; aim to spend time doing what you naturally love and are good at.
  • Periodically reassess: is this role using my strengths or do I feel pulled elsewhere?
  • Frameworks like the PM archetypes help people legitimize non-“classic” PM profiles.

Developing as a Great Product Person

• Dual requirement:
  • Obsess over craft and details.
  • Have judgment about which details actually matter to the business and users.
• Many early mistakes are over-indexing on minutiae at the expense of bigger levers (e.g., marketplace dynamics, key user jobs).
• The best PMs care deeply about the product and understand where to focus effort for impact.

User Empathy, Research & Design Thinking

• Dogfooding and empathy

  • There is no substitute for personally experiencing your product:
    • Peter drove as an Uber driver for two weeks before joining.
    • Uber CPO Sachin Kansal has done 700–800 rides as a driver.
  • Real-world context (e.g., driving 60 mph with a far-away phone vs office testing) reveals critical, non-obvious issues.

• IDEO / d.school design framework

  • Five stages of design thinking:
    1. Empathize – deeply feel user pain, not just observe it.
    2. Define – articulate the problem clearly and precisely.
    3. Ideate – generate solution ideas.
    4. Prototype – build quick approximations.
    5. Test – validate with users and data.
  • Peter emphasizes the first two stages as often undervalued:
    • Empathy requires presence and focus, not summaries.
    • Defining uses language to compress and align around the real problem.

• Caution with AI summaries in research

  • Doing interviews and then asking an LLM to summarize can remove the emotional signal.
  • You cannot empathize with a summary; you must hear tone, pauses, and nuance yourself.
  • Anecdotes can outweigh data when they reveal fundamental truths (echoing Jeff Bezos’s “if anecdote vs data conflict, trust the anecdote”).

Case Studies

News Feed (Facebook)

• Goal: long-lasting architecture for information consumption and sharing.
• Approach:
  • Designed full loop: creation, ranking, interaction (likes/comments), notification feedback.
  • Thought deeply about how humans want to consume information socially.
• Outcome:
  • Core structure has remained largely stable for over a decade.

Uber Reserve

• Problem: riders with early flights lack peace of mind if they must hope a car is available at 4 a.m.
• Insight: primary job is peace of mind, not raw ride request.
• Solution:
  • Uber Reserve – schedule rides ahead of time, often at a premium price.
  • Carefully designed flows (e.g., warnings if pickup time risks missing flight).
  • Balanced rider peace of mind with driver incentives and reliability constraints.
• Outcome:
  • Built around a simple idea with strong craft on what truly mattered.
  • Became a multibillion-dollar, high-margin business line.

Instagram Bolt (Failure Lesson)

• Product: separate, camera-first app to reduce sharing pressure and send quick photos.
• Advantages:
  • Top-tier design and performance thanks to Instagram engineering and design.
• Launch: tested in markets like New Zealand / Australia.
• Result:
  • Retention curves did not asymptote; product failed to stick.
• Learning:
  • Even elite teams with excellent taste cannot perfectly predict hits.
  • Failure is acceptable if you learn, salvage tech, and move on.
  • Re-emphasizes importance of retention and user-job fit over pure craft.

Career Choices & Company Evaluation

• Why Peter left Google for early Facebook

  • Facebook had a strong, explicit mission (“make the world more open and connected”) clearly rooted in human needs.
  • Team deeply understood loneliness, connection, and sharing as fundamental human drivers.
  • Felt more aligned with his fascination for psychology and “how humans are wired” than Google did at the time.
  • Decision also driven by optimizing for learning; saw more growth opportunities at Facebook.

• General advice when choosing roles

  • Optimize for learning above prestige or comfort.
  • Look for:
    • Mission and product that align with fundamental human behavior.
    • Clear, non-vapid articulation of purpose vs generic social site statements.
    • Founders with unique, powerful insights that teach you something new.
  • Instagram quote: “We may not be right, but at least we’re not confused.”
    • Value clear conviction over hedged ambiguity.

Decisions

• Model designer as a defined function at OpenAI

  • Decision to formalize and staff “model designer” roles around people like Joanne Jen.
  • Recognized unique combination of tech depth + taste as a key capability for ChatGPT.

• Uber Reserve investment

  • Decision to prioritize solving “peace of mind” over adding more UI features.
  • Hard commitment to simple product with strong attention to the right details.

Open Questions

• How should education systems structurally change (curriculum, assessment, teacher training) to focus on abstraction, prompts, and creativity?
• In a world where foundational models commoditize intelligence, what new moats beyond data and workflows will emerge?
• How far can product craft and workflow advantage alone carry startups against incumbents as model capabilities converge?
• What additional archetypes might exist for other functions (design, engineering) analogous to the five PM types?